GZ Lin

The companies getting this right release models openly. AI will become common open infrastructure, like Linux. The goal is adoption, not walled gardens. AI is a tool, not an ideology. We didn't build a "Microprocessor Product." We built computers. Same thing here.

The dream of always-on AI agents running on voice has a problem: people hate talking to their devices. Voice UI only works well for a simple things, alarms, weather, checking messages. For everything else? It's not private, not accurate, and always hovering to misfire. The phone stays because you can stow it, forget it exists, and pull it out when needed.

"Less clicks = more." AI agents will make UI design obsolete. This sounds utopian. It might be dystopian. Friction has its place. An "are you sure?" confirmation is friction. Often friction protects you. Sometimes it protects the company at your expense. Less isn't always more.

Apple says we never ship a technology. Google has quietly been making AI a feature, not a product: call scam detection, spam filtering, Google Lens, Magic Cue. Apple philosophy is defensible. But their execution has been cautious to the point of invisibility.

Four years post-GPT-3.5: what has materially changed for the average person? Sure, translation. Sure, small business websites. But the internet also did not have an immediately obvious daily use case until it rewrote everything. ChatGPT reached 900M users faster than any product in history. That speed tells you something.

Futurists say AI agents will book your rides, plan your trips, automate the minutiae of life. A majority of people do not actually want their lives fully automated. The things we do for ourselves, grocery lists, gift ideas, trip planning, are not problems to solve. They are part of living.

Apple does not need a killer AI product. Neither did Microsoft need a killer Windows product or any company need a killer wireless networking product. AI is infrastructure, not a thing you ship. It becomes better by being everywhere and nowhere at once.

Three classes of models in the 2026 landscape: 1) Closed frontier — strongest knowledge work, coding agents 2) Open frontier — good enough for many use cases, growing gaps in specialized domains 3) Small open models — distributed intelligence, specific tools, 100x cheaper complements We're still years away from understanding what it means to have this magnitude of intelligence served at the marginal cost of electricity. Open models are how we get there without building a new walled garden.

Open models won't win by being better closed models. They'll win by being better complements. The business isn't releasing weights. The business is building on top of them — tool integrations, deployment patterns, domain expertise. As Rosenberg said about open systems 20 years ago: 'Advantage derives from understanding the fast-moving system better than anyone else.'

Bill Gurley's analogy about Android and Chrome still applies to AI: Google gave away Android and Chrome not because they were profitable products, but because they were expensive defensive moats. They took layers between themselves and the consumer and made them free (or less than free).

Distillation — the main method for copying closed model performance — requires more creativity now. Previously you could train on the full completion. Now the important part is complex RL environments and agent prompts. These are harder to hide. And Chinese labs consistently complain about computational restrictions. The gap will stabilize or grow. Not because open labs are lazy. Because the training paradigm has shifted.

The Qwen family is already doing this. Qwen3.6-35B-A3B and Qwen3.6-27B are open-weight dense coding models marketed on general benchmarks, but they're genuinely suited for the sub-task work that frontier agents offload. The hype about open catching frontier distracts from this enormous, under-explored demand.

There's massive pressure to shift repetitive niche tasks off the best closed models onto small, open models that are 10x faster and 100x cheaper. The problem? Almost no one is building data and fine-tuning engines for economically viable tasks on the smallest models. Everyone chases leaderboard benchmarks.

In a world dominated by coding agents, I want small open models that Claude Code is desperate to use as a tool. Sub-agents unlocking entirely new areas of work. Brain-numbingly boring and specific. Not general benchmarks. Example: an open model fine-tuned specifically for parsing cloud provider API responses and extracting cost data. Fast, cheap, specialized.

I've been thinking about the wrong question for open models. Everyone asks: 'When will open models catch frontier performance?' The better question: 'What should open models be good at that frontier models aren't?' The answer isn't caught-up performance. It's something completely different.

The open vs closed question keeps getting asked. I think the better question: what layer of the stack should be open? Weights alone are never enough. We need open harnesses, open tool integrations, open deployment patterns. That's where the real public value lives — not in static model downloads.

Open models also need to think about what they're NOT. Don't try to compete on frontier benchmarks if you can't. Instead: build models that Claude Code desperately wants to use as plugins. Let its sub-agents unlock entirely new areas of work. Brain-numbingly boring and specific is a feature, not a bug.

The real open model moat isn't releasing good weights. As Google's Jonathan Rosenberg wrote in 2009 regarding Android (still applies to AI): "A competitive advantage doesn't derive from locking in customers, but from understanding the fast-moving system better than anyone else." The company winning open AI will be the one that understands systems: weights + tools + harness

The open model business landscape will bifurcate into three classes: 1) True frontier closed models → strongest knowledge work and coding agents 2) Open frontier models → best open-weight, competing on same dimensions, but with growing gaps in specialized domains 3) Small open models as distributed intelligence → 10x faster, 100x cheaper sub-task tools

Chinese open-source labs (Qwen, DeepSeek, Kimi K2, GLM-4.5) are not only catching up, they're also building different paths to value: - Alibaba Qwen3.6: open agentic coding models (35B-A3B, 27B) - Moonshot Kimi K2.6: 1T-parameter MoE, leading open coding - Z AI GLM-4.5: fused reasoning + coding + agents - Deepseek V4: Pioneering efficient inference on non NVIDIA